Elevating large language model reasoning ability with auto-enhanced zero-shot prompts

Zero-shot Chain of Thought (Zero-shot CoT) has been shown to effectively enhance the reasoning abilities of Large Language Models (LLMs). However, the prompts used in Zero-shot CoT are typically fixed or manually designed, often without accounting for the unique properties of different LLMs. Automatically designing effective prompts in zero-shot CoT remains a challenging task. In this paper, we propose an Auto-enhanced Zero-shot Prompt Strategy (AZPS) that automatically adjusts the prompts in zero-shot CoT. Specifically, we model the process of enhancing prompts as a retrieval process. During training, our AZPS first builds a set of zero-shot prompts; then, we obtain the evaluation score based on the LLMs for each zero-shot prompt in the training data, and we learn a retrieval model to select zero-shot prompts that can obtain satisfactory performance in the set based on the input questions in the training data. During testing, for one question, the retrieval model selects proper zero-shot prompts from the pre-generated set, and a majority voting strategy is adopted to further improve the reasoning ability of the LLMs. We conducted experiments using three LLMs on three mathematical reasoning datasets. The results demonstrate the effectiveness of our method compared with state-of-the-art automatic zero-shot CoT methods.